Open Pearl Miner v1.9.0 public release
What new
- Fused CuTe Ampere/Ada kernel (sm_80+). 128×256 CTA tile, in-mainloop transcript fold, multistage cp.async.cg + ldmatrix, <8,1,1> TiledMMA (each warp owns a 16-row band so the R-block fold is a direct in-register shfl_xor). Engaged automatically for the real mining shape (region multiple of 128×256, R multiple of 32). Measured 42.7 TH/s on an RTX 4050 (Ada), bit-exact with the DP4A transcript.
- Dependency-free ldm fallback. Raw-PTX mma.sync.m16n8k32 with 16 N-fragments per warp (A fragment loaded once, reused ×16) for shapes the CuTe fold doesn't cover, plus a cascade of 64×256 / 64×128 tilings.
- Pascal DP4A dual-tile (NT=2). Each warp now computes two adjacent 16×16 tiles, loading the A fragment once for both (shared-memory traffic per FLOP cut ~1.5×), with int2 global→shared staging. Bit-exact: identical dp4a inputs and order.
- Leaner bundle. Excluded stray transitive deps (numba/llvmlite, pandas, lxml, pywin32) from the frozen build — back to ~30 MB.
Experimental
- CUDA path builds clean for sm_61 / sm_86 / sm_89; the Windows DLL is smoke-tested on a Tesla P40 (Pascal DP4A, pow_split OK).
- The Pascal NT=2 change is functional (P40 smoke test OK) but transcript equality on your card is worth a first-shares check as well.
Performance
- The CuTe kernel benchmarked 42.7 TH/s bit-exact on the RTX 4050